An object pose estimation method, apparatus, device and medium

Abstract

This application discloses an object pose estimation method, apparatus, device, and medium. The method includes: fusing RGB features of an RGB image and point cloud features of a point cloud representation to obtain a first fused feature; encoding a time step to generate a time step feature, and fusing the time step feature with the first fused feature to obtain a second fused feature; inputting the second fused feature into a pre-trained diffusion model, and using the second fused feature as a guiding condition input during the back-diffusion process of the diffusion model, predicting noise in the noisy pose using the denoising network of the diffusion model, and removing noise in the noisy pose to reconstruct the target pose. This application, by fusing RGB features and point cloud features and using a diffusion model for pose reconstruction, has the advantages of achieving high-precision pose estimation without external depth input and based only on conventional RGB images, point cloud representations, and a diffusion model, thus improving the accuracy of the model's prediction results.

Description

Technical Field

[0001] This application relates to the field of pose estimation technology, specifically to an object pose estimation method, apparatus, device, and medium. Background Technology

[0002] Object pose estimation is a current research hotspot in fields such as autonomous driving and robot navigation, serving as a core means for understanding and interacting with the environment. This task aims to predict an object's 3D (Three Dimensional) translation, 3D rotation, and 3D scale relative to a camera. Early methods used 2D (Two Dimensional) / 3D constraints and geometric priors to extract the object's pose. These constraints typically required auxiliary models, pseudo-LiDAR for depth estimation, or combining image features with depth maps as model priors. While these depth sensor-based and depth prior-based methods achieved excellent performance, their over-reliance on depth information is problematic. Furthermore, high-precision, high-performance depth sensors are usually expensive and energy-intensive, and most mobile devices do not possess them. This issue directly limits the deployment flexibility of pose estimation systems in autonomous driving and robot navigation, thus affecting the reliability of environmental understanding and interaction. Summary of the Invention

[0003] This application provides a method, apparatus, device, and medium for estimating the pose of an object to solve the aforementioned technical problems.

[0004] Firstly, this application provides an object pose estimation method, comprising: Obtain the RGB image to be processed and its point cloud representation; The RGB features of the RGB image and the point cloud features of the point cloud representation are extracted respectively, and the RGB features and the point cloud features are fused to obtain the first fused feature; The current time step of the diffusion model is encoded to generate a time step feature, and the time step feature is fused with the first fusion feature to obtain a second fusion feature; The second fused feature is input into the pre-trained diffusion model. During the backward diffusion process of the diffusion model, the second fused feature is used as the guiding condition input. The noise in the noisy pose is predicted by the denoising network of the diffusion model, and the noise in the noisy pose is removed to reconstruct the target pose. The noisy pose is obtained by adding Gaussian noise to the real target pose during the forward diffusion process of the diffusion model.

[0005] In one embodiment of this application, the RGB features and the point cloud features are fused to obtain a first fused feature, including: The RGB features and the point cloud features are concatenated to obtain the concatenated features; The spliced ​​features are sequentially processed through convolution and normalization to obtain the gating weights; The first fused feature is obtained by weighting and summing the RGB features and the point cloud features using the gating weights.

[0006] In one embodiment of this application, the time step feature is fused with the first fusion feature to obtain the second fusion feature, including: Obtain a random mask variable, and process the first fused feature based on the random mask variable; The first fusion feature, after being processed by the random mask variable, is added element-wise to the time step feature to generate the second fusion feature; wherein, the probability that the first fusion feature is a null value is the first probability, and the probability that the first fusion feature is a true condition is the second probability, and the first probability and the second probability are complementary.

[0007] In one embodiment of this application, the method of predicting noise in noisy poses using the denoising network of the diffusion model includes: Perform one prediction based on the true conditions and one prediction based on null values ​​respectively to obtain the conditional prediction noise corresponding to the true conditions and the unconditional prediction noise corresponding to the null values. The conditional prediction noise and the unconditional prediction noise are weighted using a preset guiding weight hyperparameter to obtain the noise in the noisy posture.

[0008] In one embodiment of this application, the method of predicting noise in noisy poses using the denoising network of the diffusion model includes: The noisy input of the diffusion model is encoded using a multilayer perceptron to obtain pose features; The posture features are concatenated with the guiding condition input to obtain the network input features of the denoising network; The network input features are processed by the denoising network to obtain effective features; The effective features are added element by element to the network input features; Based on the features obtained by element-wise addition and the noise in the noisy posture output by the multilayer perceptron.

[0009] In one embodiment of this application, the denoising network includes at least one backdiffusion block, the backdiffusion block including residual units and self-attention units; processing the network input features through the denoising network includes: The residual unit performs feature transformation on the network input features; The global features are obtained by capturing the global dependencies between different positions in the features after feature transformation through the self-attention unit.

[0010] In one embodiment of this application, obtaining the point cloud representation of the RGB image includes: A coarse depth estimation is performed on the RGB image to obtain a depth map; The RGB image is segmented into instances to obtain a binary mask; The depth map is cropped using the binary mask, and the cropped depth map is converted into a point cloud representation through back projection.

[0011] Secondly, this application provides an object pose estimation device, comprising: The data acquisition module is used to acquire the RGB image to be processed and the point cloud representation of the RGB image; The first feature fusion module is used to extract the RGB features of the RGB image and the point cloud features of the point cloud representation, and fuse the RGB features and the point cloud features to obtain the first fused feature; The second feature fusion module is used to encode the current time step of the diffusion model to generate time step features, and fuse the time step features with the first fusion feature to obtain the second fusion feature; The pose reconstruction module is used to input the second fused feature into a pre-trained diffusion model. During the backward diffusion process of the diffusion model, the second fused feature is used as a guiding condition input. The noise in the noisy pose is predicted by the denoising network of the diffusion model, and the noise in the noisy pose is removed to reconstruct the target pose. The noisy pose is obtained by adding Gaussian noise to the real target pose during the forward diffusion process of the diffusion model.

[0012] Thirdly, the electronic device provided in this application includes: One or more processors; A storage device for storing one or more programs that, when executed by one or more processors, enable an electronic device to implement an object pose estimation method.

[0013] Fourthly, this application provides a computer-readable storage medium storing a computer program thereon, which, when executed by a computer's processor, causes the computer to perform an object pose estimation method.

[0014] The beneficial effects of this application are as follows: The object pose estimation method of this application acquires an RGB image to be processed and a point cloud representation of the RGB image; extracts the RGB features of the RGB image and the point cloud features of the point cloud representation respectively, and fuses the RGB features and the point cloud features to obtain a first fused feature; encodes the current time step of the diffusion model to generate a time step feature, and fuses the time step feature with the first fused feature to obtain a second fused feature; inputs the second fused feature into a pre-trained diffusion model, and during the backward diffusion process of the diffusion model, uses the second fused feature as a guiding condition input, uses the denoising network of the diffusion model to predict the noise in the noisy pose, and removes the noise in the noisy pose to reconstruct the target pose; wherein, the noisy pose is obtained by adding Gaussian noise to the real target pose during the forward diffusion process of the diffusion model. This application only requires acquiring conventional RGB images and generating corresponding point cloud representations when reconstructing the target pose, without relying on external depth sensors, thus avoiding additional depth acquisition hardware and significantly reducing system costs. On this basis, by fusing RGB features and point cloud features, the denoising network can simultaneously utilize the color information of the image and the geometric structure information of the point cloud while gradually removing noisy poses. Furthermore, with the help of the data augmentation strategy inherent in the diffusion model, it can learn more discriminative feature representations, ultimately achieving higher accuracy target pose reconstruction.

[0015] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0016] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application. It is obvious that the drawings described below are merely some embodiments of this application, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort. In the drawings: The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application. It is obvious that the drawings described below are merely some embodiments of this application, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0017] In the attached diagram: Figure 1 This is a schematic diagram of an object pose estimation method according to an embodiment of this application; Figure 2This is a flowchart illustrating an object pose estimation method according to another embodiment of this application; Figure 3 This is a schematic diagram illustrating the process of fusing RGB features and point cloud features according to an embodiment of this application; Figure 4 This is a schematic diagram of the process of using a diffusion model-based denoising network to predict noise in noisy poses according to an embodiment of this application. Figure 5 This is a schematic diagram illustrating the process of obtaining a point cloud representation of an RGB image according to an embodiment of this application; Figure 6 This is a schematic block diagram of an object pose estimation device according to an embodiment of this application; Figure 7 A schematic diagram of a computer system suitable for implementing the memory of the embodiments of this application is shown. Detailed Implementation

[0018] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. This application can also be implemented or applied through other different specific embodiments. Various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. In the absence of conflict, the following embodiments and features in the embodiments can be combined with each other.

[0019] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of this application. The drawings only show the components related to this application and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the shape, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0020] In the following description, numerous details are explored to provide a more thorough explanation of embodiments of the present application. However, it will be apparent to those skilled in the art that embodiments of the present application may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring embodiments of the present application.

[0021] Object pose estimation is a current research hotspot in fields such as autonomous driving and robot navigation, serving as a core means for understanding and interacting with the environment. This task aims to predict the 3D translation, 3D rotation, and 3D scale of an object relative to a camera. Early methods used 2D / 3D constraints and geometric priors to extract the object's pose. These constraints typically required auxiliary models, pseudo-LiDAR for depth estimation, or combining image features with depth maps as model priors. These depth sensor-based and depth prior-based methods achieved excellent performance, but their dependence on depth information was too high. However, the integration of depth sensors increased system hardware costs and power consumption, and most mobile devices lack such sensors, making depth-based methods impractical in resource-constrained environments. This problem directly limits the deployment flexibility of pose estimation systems in autonomous driving, robot navigation, and other fields, thus affecting the reliability of environmental understanding and interaction.

[0022] Based on this, this application provides an object pose estimation method to address the problem that the lack of a physical depth sensor makes it difficult to effectively extract geometric priors and achieve real-time pose estimation for unknown objects. Please refer to [link to relevant documentation]. Figure 1 , Figure 1 This is a flowchart illustrating an object pose estimation method according to an embodiment of this application. Figure 1 In this context, DINOv2 represents the depth estimation network, SOLOv2 represents the instance segmentation network, ResNet18 represents the RGB feature extraction network, Back-Projection represents back projection, Conv represents a convolutional layer, Sigmoid represents an activation function, Gate represents gate weights, PointNet represents a point cloud feature extraction network, MLP represents a multilayer perceptron, Forward Diffusion represents the forward diffusion process, Resblock represents a residual unit, and Self-Attention represents a self-attention unit. Figure 1 In this method, the following steps are taken: 1) Obtain the RGB image to be processed and its point cloud representation; 2) Extract the RGB features of the RGB image and the point cloud features of the point cloud representation, and fuse the RGB features and point cloud features to obtain the first fused feature; 3) Encode the current time step of the diffusion model to generate time step features, and fuse the time step features with the first fused feature to obtain the second fused feature; 4) Input the second fused feature into the pre-trained diffusion model, and use the second fused feature as the guiding condition input during the backward diffusion process of the diffusion model to predict the noise in the noisy pose and remove the noise in the noisy pose to reconstruct the target pose; 5) The noisy pose is obtained by adding Gaussian noise to the real target pose during the forward diffusion process of the diffusion model.

[0023] For ease of understanding, the following explains some key terms in this embodiment: RGB (Red, Green, Blue) images: These images represent pixel color information using a combination of red, green, and blue color channels. This format is commonly used as input data in computer vision, providing texture and color details.

[0024] Point cloud: A dataset consisting of a series of points in three-dimensional space. Each point typically contains its three-dimensional coordinates and possibly additional information such as color and intensity. Point clouds can directly represent the three-dimensional geometric information and spatial distribution information of an object.

[0025] RGB features: Vectors extracted from RGB images that can represent image content and semantic information.

[0026] Point cloud features: Vectors extracted from the point cloud representation that characterize the geometric structure and spatial distribution of the point cloud.

[0027] Time step features: Feature representations obtained after encoding the time step.

[0028] Random mask variable: Introduces randomness during training to control the availability of conditional information. The random mask variable can be a discrete binary variable, such as 0 or 1, where 0 indicates the conditional information is empty and 1 indicates it is retained. The random mask variable follows a Bernoulli distribution m ~ Bernoulli (0.9) and can be randomly sampled from the Bernoulli distribution.

[0029] Diffusion Model: A generative model consisting of two stages, each represented by a Markov chain. The first stage is the forward diffusion stage, which progressively adds Gaussian noise with predefined mean and variance to the true target pose; the second stage is the backward diffusion stage, which uses a neural network to progressively remove the Gaussian noise from the pose and reconstruct the target pose.

[0030] The forward diffusion stage is used to construct an approximate posterior distribution. ,here x 0 represents the true target pose. p 0. This process uses Markov chain modeling. Over T time steps, Gaussian noise is gradually added to the true target pose p0 (the initial noiseless pose), thus generating a series of noisy poses p1… p T This refers to a noisy input, where the addition of noise is controlled by cosine modulation, defining a set of increasing noise variance coefficients. Its satisfaction The entire forward process can be decomposed into the product of conditional probabilities at each step:

[0031] The transition formula for each step is as follows:

[0032] in, This indicates that during the diffusion process, given an initial state... x Under condition 0, the sequence of all states x1, x2, ..., x from time 1 to T. T The probability distribution; Let represent the noise variance coefficient at step t, where t represents the current time step in the diffusion process, and T represents the total number of time steps in the diffusion process. p t This represents the noisy pose (noisy input) at time step t. Indicates the first pose at each time step This represents standard Gaussian noise randomly sampled at time step t.

[0033] make , , This represents the pose at the previous time step during the diffusion process at step t. The retention ratio, This represents the overall retention rate from step t to the final step. The larger the value, the less noise is added. The above formula can be transformed into sampling the degenerate pose at any given time step:

[0034] in, This indicates that during the forward diffusion process, given the initial true pose... p Under the condition of 0, the noisy pose at step t p t The conditional probability distribution of follows a Gaussian distribution. Represents the mean of the distribution. The covariance matrix of the distribution is represented. I Represents the identity matrix.

[0035] The reverse diffusion stage under real conditions Guided by standard Gaussian noise posture The actual target pose is recovered. However, since the distribution of the entire data is unknown, it is difficult to directly predict the conditional probability. Based on the properties of the diffusion model, assuming the noise is Gaussian noise, the conditional probability of the reverse process is... It itself approximates a Gaussian distribution, and the Gaussian distribution is mainly determined by two parameters, namely the mean. and variance Therefore, a denoising model is learned during the reverse process. To predict the mean and variance, the network uses the current noisy pose. p t Time step t and condition information f c For input, output:

[0036] Denoising network: It is part of the diffusion model and is trained to predict noise in a given noisy data.

[0037] Feature splicing: Connecting different features along a specific dimension to form a comprehensive feature representation that contains information from at least two modalities.

[0038] Multi-Layer Perceptron (MLP): A type of feedforward neural network consisting of at least three layers: an input layer, a hidden layer, and an output layer. It mainly learns complex patterns and features by performing multi-layer transformations on the input data through non-linear activation functions.

[0039] Guiding condition input: Auxiliary information used to guide the denoising network in noise prediction during the back diffusion process of the diffusion model, including but not limited to object category information, scene context information, or user-specified pose constraints.

[0040] Residual unit: A widely used structure in deep neural networks that allows information to be passed directly from the previous layer to the next by introducing skip connections, thereby alleviating the gradient vanishing problem in deep network training and enabling the network to learn deeper feature representations.

[0041] Self-attention units (SAUs) are mechanisms that can compute correlations between different locations in an input sequence or feature map. They capture global dependencies between different parts of a feature by generating query, key, and value vectors for each location, computing the dot product of the query and all keys to obtain attention weights, and then applying these weights to a weighted sum of the value vectors.

[0042] Feature transformation: By using a series of mathematical operations (such as convolution, pooling, activation functions, etc.) to change the representation of input features, higher-level, more abstract or more discriminative information can be extracted, thus mapping the original input data to a new feature space.

[0043] Global dependencies: Interactions exist between different locations in a feature map. In pose estimation, for example, the pose of one object may be strongly correlated with the pose of another object.

[0044] Please see Figure 2, Figure 2 This is a flowchart illustrating an object pose estimation method according to an embodiment of this application. The object pose estimation method includes at least steps S210 to S240, which are described in detail below: Step S210: Obtain the RGB image to be processed and the point cloud representation of the RGB image; For example, RGB images can be obtained directly through a camera device. The point cloud representation can be obtained by performing depth estimation on the RGB image to obtain a depth map; performing instance segmentation on the RGB image to obtain a binary mask; cropping the depth map using the binary mask; and then converting the cropped depth map into a point cloud representation through backprojection.

[0045] Step S220: Extract the RGB features of the RGB image and the point cloud features of the point cloud representation, and fuse the RGB features and the point cloud features to obtain the first fused feature; For example, RGB features can be extracted from images using pre-trained deep learning models such as convolutional neural networks to capture local and global visual patterns. Point cloud features can be extracted from point cloud data using specialized point cloud processing networks to capture the geometric structure and spatial distribution of objects.

[0046] Step S230: Encode the current time step of the diffusion model to generate time step features, and fuse the time step features with the first fusion feature to obtain the second fusion feature; For example, time steps can be encoded using positional encoding, which maps integer time steps to high-dimensional continuous vectors, enabling them to carry time information.

[0047] Step S240: The second fusion feature is input into the pre-trained diffusion model. During the backward diffusion process of the diffusion model, the second fusion feature is used as the guiding condition input. The noise in the noisy pose is predicted by the denoising network of the diffusion model, and the noise in the noisy pose is removed to reconstruct the target pose. The noisy pose is obtained by adding Gaussian noise to the real target pose during the forward diffusion process of the diffusion model.

[0048] As a specific implementation, when an object enters the camera's field of view, its RGB image is acquired, and then a point cloud representation of the object is generated based on the RGB image. These acquired RGB images and point cloud representations are then fed into a first feature fusion module. In this module, the RGB images are input into a pre-trained convolutional neural network to extract RGB features representing visual information such as color, texture, and shape. Simultaneously, the point cloud representation is input into a point cloud processing network to extract point cloud features representing the object's three-dimensional geometric structure and spatial distribution. Next, the RGB features and point cloud features are concatenated using feature vectors to form a first fused feature. This first fused feature integrates the object's two-dimensional appearance and three-dimensional structural information, providing a description for subsequent pose estimation. The first fused feature is then fed into a second feature fusion module. Simultaneously, a time-step feature is also input into the second feature fusion module and fused with the first fused feature to obtain a second fused feature containing temporal information and multimodal information of the object, which serves as a guiding condition input for pose reconstruction using a diffusion model. Subsequently, the second fused feature is input into a pre-trained diffusion model. In the back-diffusion process of the diffusion model, the second fusion feature serves as a guiding condition input, instructing the denoising network of the diffusion model to predict the noise present in the noisy pose based on the current noisy pose. The predicted noise is then subtracted from the noisy pose, resulting in a less noisy pose. This denoising process iterates for multiple time steps. At each time step, the denoising network predicts and removes noise based on the current noisy pose and the guiding condition input. Through this stepwise denoising approach, the target pose of the object is finally reconstructed from the initial noisy pose.

[0049] In one embodiment, point cloud feature extraction can be achieved using PointNet, which inputs point cloud data and obtains its global feature vector as the point cloud features. RGB feature extraction can be achieved using a ResNet network, which inputs an RGB image and obtains its deep feature map as the RGB features.

[0050] For example, point cloud feature extraction based on the PointNet network includes: Step 1, acquiring raw point cloud data, which contains N points, each with three-dimensional spatial coordinates; Step 2, spatially aligning the raw point cloud data using a first transformation network to obtain an aligned point cloud; Step 3, independently inputting each point in the aligned point cloud into a multilayer perceptron with shared weights, mapping the coordinates of each point to a high-dimensional feature vector, and obtaining N point-by-point feature vectors; Step 4, performing feature space alignment on the point-by-point feature vectors using a second transformation network; Step 5, performing max pooling on the point dimension of the aligned point-by-point feature vectors, taking the maximum value on each feature channel, and outputting a first global feature vector, which is used to characterize the point cloud features of the entire point cloud.

[0051] For example, extracting RGB features based on a ResNet network includes: Step 1, acquiring RGB image data, the image data having height H, width W, and 3 color channels; Step 2, sequentially performing initial convolution, batch normalization, ReLU activation, and max pooling on the RGB image data to obtain an initial feature map; Step 3, sequentially inputting the initial feature map into multiple residual modules, each residual module containing at least two convolutional layers and one skip connection, the skip connection directly adding the input of the residual module to the output of the convolutional layer to obtain an output feature map; after passing through multiple residual modules, a final feature map is obtained; Step 4, performing global average pooling on the final feature map, taking the average value in the spatial dimension, and outputting a second global feature vector, the second global feature vector being used to characterize the RGB features of the RGB image.

[0052] Please see Figure 3 , Figure 3 This is a schematic diagram of a process for fusing RGB features and point cloud features according to an embodiment of this application. The process of fusing RGB features and point cloud features to obtain a first fused feature includes at least steps S310 to S330, which are described in detail below: Step S310: The RGB features and point cloud features are concatenated to obtain the concatenated features; For example, the stitching of RGB features and point cloud features can be stitching in the channel dimension or stitching in the spatial dimension.

[0053] Step S320: The concatenated features are sequentially processed through convolution and normalization to obtain the gating weights; The convolution operation is implemented using convolutional layers, which learn key information from the concatenated features for calculating the gating weights. These convolutional layers can be one-dimensional or two-dimensional and can contain multiple kernels to extract features from different patterns. Normalization is achieved using sigmoid activation, a non-linear function whose output value ranges from 0 to 1, used to generate gating signals as weights.

[0054] Step S330: The RGB features and point cloud features are weighted and summed using gating weights to obtain the first fused feature.

[0055] As a specific implementation method, when performing feature fusion, the RGB features are first... and point cloud features The features are concatenated to form a comprehensive feature representation that includes information from both modalities. The concatenated features are input into a convolutional layer, and deep patterns are learned and extracted from the concatenated features to evaluate the importance of each modality. The Sigmoid activation function maps the output of the convolutional layer to the range of 0 to 1, thereby generating a set of gating weights. Finally, the original RGB features and point cloud features are weighted and summed using gating weights to obtain the first fused feature. .

[0056] The dynamic weighted fusion method proposed in this application can intelligently adjust the contribution of different modal features according to the characteristics of the input features, which avoids redundancy and noise problems that may be caused by simple splicing to a certain extent, thereby ensuring the quality and discriminative power of the fused features.

[0057] In some embodiments described above in this application, a second fusion feature is obtained by fusing time-step features with a first fusion feature, and this second fusion feature is used as a guiding condition for the diffusion model. However, during the training phase of the diffusion model, if the model is always guided by complete real conditions (i.e., the first fusion feature), the extended model may become overly reliant on this conditional information. This could lead to an impact on the denoising ability and pose reconstruction accuracy of the extended model during the inference phase when the actual conditional information (such as RGB images or point cloud representations) is noisy or missing, thereby reducing the model's generalization ability and robustness.

[0058] Based on this, the time step features are fused with the first fusion feature to obtain the second fusion feature, including: obtaining a random mask variable and processing the first fusion feature based on the random mask variable; adding the first fusion feature processed by the random mask variable to the time step features element by element to generate the second fusion feature; wherein, the probability that the first fusion feature is a null value is the first probability, the probability that the first fusion feature is a true condition is the second probability, and the first probability and the second probability are complementary.

[0059] As a specific implementation method, after obtaining RGB features and point cloud features and fusing them to obtain the first fused feature, the random mask variable is first obtained. m Randomness is introduced during training by using a random mask variable to control the availability of conditional information. Then, the first fused feature is processed based on the random mask variable to obtain the final feature. The first fused feature is selectively replaced based on a random mask variable to simulate different levels of conditional guidance. When the random mask variable indicates that the conditional information is masked, the first fused feature can be replaced with a preset zero vector, i.e., a null value; when the random mask variable indicates that the conditional information is retained, the first fused feature remains unchanged, i.e., the first fused feature is retained as a true condition. The probability that the first fused feature is replaced with a zero vector (null value) is the first probability, and the probability that the first fused feature is retained as a true condition is the second probability. The first and second probabilities are complementary, i.e., the sum of the first and second probabilities is 1. As a specific implementation, the first probability can be set to 0.1, and the second probability can be set to 0.9. Finally, the first fused feature processed by the random mask variable is compared with the time step feature. Element-wise addition is performed to superimpose the values ​​of the two feature vectors at corresponding positions, thereby achieving information fusion between the first fused feature and the time step feature to generate the second fused feature. .

[0060] In this application, the above-mentioned method can learn how to effectively use complete conditional information to predict noise and reconstruct pose, and can also learn how to complete the denoising task by relying only on time step information and its own internal representation when conditional information is missing.

[0061] In one embodiment, time step features The time step T can be encoded using a multilayer perceptron (MLP), which is a vector with a dimension of 1×512.

[0062] In one embodiment, the noise in a noisy pose is predicted using a denoising network based on a diffusion model, including: performing a prediction based on true conditions and a prediction based on null values, respectively, to obtain conditional prediction noise corresponding to true conditions and unconditional prediction noise corresponding to null values; and weighting the conditional prediction noise and unconditional prediction noise using a preset guiding weight hyperparameter to obtain the noise in the noisy pose.

[0063] As a specific implementation, during the back-diffusion process of the diffusion model, the denoising network performs two independent noise predictions for the noisy pose to be processed. The first is a conditional noise prediction guided by the ground truth conditions (i.e., the second fusion feature, which includes RGB features, point cloud features, and time step features), resulting in conditional noise prediction. The second is an unconditional noise prediction based on null conditions (i.e., without providing any specific guiding information), resulting in unconditional noise prediction. After obtaining the two different predicted noises, conditional and unconditional, an adjustable guiding weight hyperparameter controls the degree of influence of the ground truth conditions on the final noise prediction.

[0064] In one embodiment, the weighting formula can be expressed as: , This indicates that noise can be conditionally predicted. Indicates unconditional predictive noise , This represents the guiding weight hyperparameter. This represents the noise in a noisy posture.

[0065] This application separates conditionally guided predictions from unconditionally guided predictions by utilizing the conditional information learned by the model during the training phase. Furthermore, it achieves fine control over the guidance intensity through adjustable guidance weight hyperparameters. This allows the model to maintain the use of conditional information while avoiding over-reliance on potentially noisy or uncertain conditional inputs, thereby obtaining more stable and accurate noise predictions during pose reconstruction.

[0066] Please see Figure 4 , Figure 4 This is a schematic flowchart illustrating the prediction of noise in a noisy pose using a denoising network based on a diffusion model, according to an embodiment of this application. The prediction of noise in a noisy pose using a denoising network based on a diffusion model includes at least steps S410 to S450, which are detailed below: Step S410: Encode the noisy input of the diffusion model using a multilayer perceptron to obtain pose features; For example, pose features include, but are not limited to, position, orientation, and scale. These features encode the noisy input of the diffusion model using a multilayer perceptron, transforming the original pose representation into features with higher dimensions and stronger semantic information.

[0067] Step S420: Concatenate the pose features with the guiding condition input to obtain the network input features of the denoising network; The concatenation operation combines pose features with guiding condition inputs to form the network input features for the denoising network. This can be achieved by directly connecting along the feature dimensions (e.g., channel dimensions or feature vector dimensions).

[0068] Step S430: The network input features are processed by a denoising network to obtain effective features; The denoising network is a core component of the diffusion model. Its function is to receive noisy input and guiding conditions, and predict the noise contained in the noisy input. The denoising network can be a deep neural network capable of learning complex noise patterns and separating them from noisy poses.

[0069] Step S440: Add the effective features to the network input features element by element; Element-wise addition is used to combine the features processed by the denoising network with the network input features of the denoising network. Specifically, it means adding the values ​​at corresponding positions in the two features to obtain a new feature.

[0070] Step S450: Based on the features obtained by element-wise addition and the noise in the noisy pose output by the multilayer perceptron.

[0071] The noise in the noisy pose is Gaussian noise superimposed on the true pose, which is ultimately predicted by the denoising network. This predicted noise will be used to subtract from the noisy pose, thereby gradually reconstructing the target pose.

[0072] As a specific implementation, the noisy input of the diffusion model is the noisy pose at the current time step t. p T The noisy pose is then fed into a multilayer perceptron (MLP) for encoding. This encoding process transforms the original noisy pose into more expressive pose features. f p This effectively extracted the core information of the pose and prepared for subsequent feature fusion. Subsequently, this pose feature... f p With pre-generated boot condition input f c (i.e., the second fusion feature) is concatenated, organically combining the pose features that need to be denoised with the rich contextual information used to guide the denoising process, forming a comprehensive network input feature for the denoising network. This denoising network, when performing noise prediction, can not only perceive the current noisy pose but also fully utilize visual, geometric, and temporal information fused from multimodal sources, thus avoiding prediction biases that may result from relying on a single information source. Next, the network input features are fed into the denoising network for processing. The denoising network learns and identifies noise patterns in the noisy pose. After processing by the denoising network, the output features are element-wise added to the original network input features. Finally, the added features are combined with the multilayer perceptron to output the noise in the noisy pose.

[0073] The above methods can more effectively utilize multi-source information to guide noise prediction, significantly improving the denoising capability and reconstruction accuracy of diffusion models in object pose estimation tasks.

[0074] In some embodiments described above in this application, a denoising network based on a diffusion model is used to predict noise in noisy poses. However, in practical applications, if the structure of the denoising network is too simple, it may be difficult to effectively capture subtle features and long-range dependencies in complex pose data, thereby affecting the accuracy of noise prediction and leading to insufficient accuracy in target pose reconstruction.

[0075] Based on this, this application proposes a specific denoising network structure. The denoising network includes at least one backdiffusion block, which includes residual units and self-attention units. The network input features are processed through the denoising network, including: performing feature transformation on the network input features through the residual units; and capturing the global dependencies between different positions in the feature transformation features through the self-attention units to obtain global features.

[0076] As a specific implementation method, backpropagation aims to learn a set of rules by... fc Conditional neural networks for estimating pose noise Then, the diffusion model's scheduler is used to denoise and recover the target pose. When the input features are fed into the denoising network, they first pass through residual units, where convolutional layers and non-linear activation functions perform preliminary feature transformation. This transformation not only extracts local features but also, due to the presence of residual connections, effectively avoids information loss or gradient vanishing problems common in deep networks, ensuring the integrity and effective transmission of feature information. Subsequently, the transformed features are fed into self-attention units, which dynamically capture the global dependencies between different locations in the feature map by calculating attention weights at those locations. This captures the correlation between each point in the feature map and other points, thereby generating a global feature containing global contextual information.

[0077] The method described in this application enables the denoising network to not only process local details but also understand the overall structure of the pose and the interactions between its parts, thereby more accurately predicting noise in noisy poses and improving the accuracy of noise prediction.

[0078] In one embodiment, the denoising network includes at least one reverse diffusion block (RDB), specifically five, with the five reverse diffusion blocks connected in sequence.

[0079] In one embodiment, due to the real noise added during the forward diffusion process Since the distribution and sampled values ​​are known, the noise reduction model can be trained using supervised learning. The loss function is defined as the mean squared error between the predicted noise and the actual noise, summed over all time steps:

[0080] in: This indicates that at time step t, it is added to the true pose representation. p Noise in 0; This represents the noise predicted by the denoising network, with the input being the guiding condition input.f c and posture features f p ; This represents the square of the L2 norm.

[0081] Please see Figure 5 , Figure 5 This is a schematic flowchart illustrating the process of obtaining a point cloud representation of an RGB image according to an embodiment of this application. The process of obtaining a point cloud representation of an RGB image includes at least steps S510 to S550, which are described in detail below: Step S510: Perform coarse depth estimation on the RGB image to obtain a depth map; For example, coarse depth estimation of RGB images can be achieved using the DINOv2 model, a self-supervised visual pre-trained model proposed by MetaAI (formerly Facebook AI Research) and an upgraded version of the DINO method. It does not rely on manual annotation and can learn general visual features from massive amounts of images, making it widely applicable to downstream tasks such as image classification, segmentation, depth estimation, and correspondence matching. The steps for coarse depth estimation of RGB images using the DINOv2 model are as follows: Step 1: Obtain the input RGB image and adjust it to a normalized image that conforms to the input size of the DINOv2 model; Step 2: Feed the normalized image into the pre-trained DINOv2 model, extract the feature map of at least one intermediate layer or the output features of all Transformer blocks to obtain a multi-scale feature representation; Step 3: Extract attention weights from the multi-head self-attention layer of the DINOv2 model, and aggregate the attention weights of different heads to generate an attention similarity matrix; Step 4: Fuse the multi-scale feature representation and the attention similarity matrix, and input them into a lightweight convolutional decoder or MLP regression head; Step 5: The decoder outputs a single-channel depth map with the same spatial resolution as the input image or after downsampling, where the depth value represents the relative depth. This depth map is the coarse depth estimation result of the original RGB image.

[0082] Step S520: Perform instance segmentation on the RGB image to obtain a binary mask; For example, instance segmentation of RGB images can be achieved using the SOLOv2 (Segmenting Objects by Locations version 2) model, which is a single-stage, anchor-free instance segmentation model. The steps for instance segmenting RGB images using the SOLOv2 model include: Step 1: Obtain the input RGB image and resize it to a preset input size to obtain a preprocessed image; Step 2: Input the preprocessed image into a backbone network (such as ResNet or ResNeXt) and a Feature Pyramid Network (FPN) to extract multi-scale feature maps, denoted as {F1, F2, ..., FL}, where L is the number of pyramid layers; Step 3: For each scale feature map Fl, divide it into S×S grid cells, each grid cell being responsible for predicting the target instance located at that position; Step 4: Input the feature map Fl into two parallel branches: a class branch and a convolutional kernel branch: the class branch outputs an S×S×C tensor through several convolutional layers, where C is the number of classes, and this tensor represents each grid cell. The probability of belonging to each category; Convolutional kernel branch: Output an S×S×E tensor through several convolutional layers, where E is the length of the parameter vector of the dynamic convolutional kernel. This tensor predicts a set of dynamic convolutional kernel parameters for each grid cell; Step 5: For each grid cell (i,j), reconstruct its corresponding depthwise convolutional kernel parameters into a 1×1×E convolutional kernel, and perform depthwise convolution with the feature map Fl to generate the original instance mask map corresponding to the grid cell; Step 6: Apply the Sigmoid function to activate each original instance mask map to obtain the instance mask probability map of each grid cell; Step 7: Multiply the category prediction probability of each grid cell with the corresponding instance mask probability map to obtain the instance segmentation result with category label, and merge the outputs of all grid cells; Step 8: Use Matrix NMS to post-process the merged instance mask, remove redundant detection, and output the final instance segmentation result (including the category label, confidence score, and binary mask of each instance).

[0083] Step S530: Use a binary mask to crop the depth map, and convert the cropped depth map into a point cloud representation through back projection.

[0084] In this step, a binary mask is used to crop the depth map in order to accurately extract the depth information of the target object from the complete depth map and remove background interference.

[0085] As a specific implementation, this application first performs coarse depth estimation on the input RGB image to infer the distance information from each pixel in the scene to the camera from the 2D image, forming a depth map to provide a foundation for subsequent point cloud generation. Next, instance segmentation is performed on the same RGB image to identify the boundaries of each object in the image and separate them from the background, generating a binary mask that indicates the pixel position of the target object in the image. Subsequently, this binary mask is used to crop the previously obtained depth map by performing a pixel-by-pixel logical AND operation between the depth map and the binary mask, retaining the depth values ​​within the mask area while setting the depth values ​​outside the mask area to invalid or zero, thereby effectively filtering out background depth information and retaining only the depth data of the target object. Finally, using the camera intrinsic and extrinsic parameter matrices, the cropped target object depth map is converted into a point cloud representation in 3D space for each pixel and its depth value using an inverse projection formula.

[0086] This application combines depth estimation and instance segmentation to effectively separate target objects from complex scenes and extract their accurate depth information, generating high-quality point clouds without background interference. This improves the efficiency and accuracy of object pose estimation methods in utilizing the 3D geometric information of input data, and avoids pose estimation errors caused by inaccurate point cloud representation or excessive background noise.

[0087] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0088] like Figure 6 As shown, this application provides a principle block diagram of an object pose estimation device. Figure 6 As shown, the object pose estimation device includes: The data acquisition module 610 is used to acquire the RGB image to be processed and the point cloud representation of the RGB image; The first feature fusion module 620 is used to extract the RGB features of the RGB image and the point cloud features of the point cloud representation, and fuse the RGB features and the point cloud features to obtain the first fused feature; The second feature fusion module 630 is used to encode the current time step of the diffusion model to generate time step features, and fuse the time step features with the first fusion feature to obtain the second fusion feature; The pose reconstruction module 640 is used to input the second fused feature into the pre-trained diffusion model. During the backward diffusion process of the diffusion model, the second fused feature is used as the guiding condition input. The denoising network of the diffusion model is used to predict the noise in the noisy pose and remove the noise in the noisy pose to reconstruct the target pose. The noisy pose is obtained by adding Gaussian noise to the real target pose during the forward diffusion process of the diffusion model.

[0089] In one embodiment, the first feature fusion module is used to concatenate RGB features and point cloud features to obtain concatenated features; the concatenated features are then subjected to convolution and normalization operations to obtain gating weights; and the RGB features and point cloud features are weighted and summed using the gating weights to obtain the first fused features.

[0090] In one embodiment, the second feature fusion module is used to obtain a random mask variable and process the first fused feature based on the random mask variable; the first fused feature after processing by the random mask variable is added element by element to the time step feature to generate the second fused feature; wherein, the probability that the first fused feature is a null value is the first probability, the probability that the first fused feature is a true condition is the second probability, and the first probability and the second probability are complementary.

[0091] In one embodiment, the attitude reconstruction module is used to perform a prediction based on the true conditions and a prediction based on null values, respectively, to obtain conditional prediction noise corresponding to the true conditions and unconditional prediction noise corresponding to the null values; the conditional prediction noise and unconditional prediction noise are weighted using a preset guiding weight hyperparameter to obtain the noise in the noisy attitude.

[0092] In one embodiment, the pose reconstruction module uses a multilayer perceptron to encode the noisy input of the diffusion model to obtain pose features; the pose features are concatenated with the guiding condition input to obtain the network input features of the denoising network; the network input features are processed by the denoising network to obtain effective features; the effective features are added element-wise to the network input features; and the noise in the noisy pose output by the multilayer perceptron is based on the features obtained by element-wise addition.

[0093] In one embodiment, the denoising network includes at least one backdiffusion block, which includes a residual unit and a self-attention unit. Processing the network input features through the denoising network includes: performing feature transformation on the network input features through the residual unit; and capturing the global dependencies between different positions in the feature transformation features through the self-attention unit to obtain global features.

[0094] In one embodiment, the data acquisition module performs coarse depth estimation on the RGB image to obtain a depth map; performs instance segmentation on the RGB image to obtain a binary mask; uses the binary mask to crop the depth map, and converts the cropped depth map into a point cloud representation through back projection.

[0095] It should be noted that the object pose estimation device and the object pose estimation method provided in the above embodiments belong to the same concept. The specific operation methods of each module and unit have been described in detail in the method embodiments and will not be repeated here. In practical applications, the object pose estimation device provided in the above embodiments can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. This is not a limitation here.

[0096] Embodiments of this application also provide an electronic device, including: one or more processors; and a memory for storing one or more programs, which, when executed by one or more processors, cause the memory to implement the object pose estimation method in the above embodiments.

[0097] Embodiments of this application also provide one or more machine-readable media having instructions stored thereon that, when executed by one or more processors, cause the processors to perform the object pose estimation method in the above embodiments.

[0098] Figure 7 A schematic diagram of a computer system suitable for implementing the memory of embodiments of this application is shown. It should be noted that... Figure 7 The computer system with the memory shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this application.

[0099] like Figure 6 As shown, the computer system 700 includes a Central Processing Unit (CPU) 701, which can perform various appropriate actions and processes, such as executing the methods described in the above embodiments, based on a program stored in Read-Only Memory (ROM) 702 or a program loaded from storage into Random Access Memory (RAM) 703. The RAM also stores various programs and data required for system operation. The CPU 701, ROM 702, and RAM 703 are interconnected via a bus 704. An Input / Output (I / O) interface 705 is also connected to the bus 704.

[0100] The following components are connected to I / O interface 705: an input section 706 including a keyboard, mouse, etc.; an output section 707 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and speakers, etc.; a storage section 708 including a hard disk, etc.; and a communication section 709 including a network interface card such as a LAN (Local Area Network) card, modem, etc. The communication section 709 performs communication processing via a network such as the Internet. A drive 710 is also connected to I / O interface 705 as needed. A removable medium 711, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on drive 710 as needed so that computer programs read from it can be installed into storage section 708 as needed.

[0101] Specifically, according to embodiments of this application, the processes described in the above-referenced flowchart can be implemented as computer software programs. For example, embodiments of this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program including a computer program for performing the object pose estimation method of the aforementioned embodiments. In such embodiments, the computer program can be downloaded and installed from a network via a communication component, and / or installed from a removable medium 711. When the computer program is executed by the central processing unit (CPU) 701, it performs various functions defined in the system of this application.

[0102] It should be noted that the computer-readable medium shown in the embodiments of this application can be a computer-readable signal medium or a computer-readable storage medium, or any combination of the two. A computer-readable storage medium can be, for example, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM) 703, read-only memory (ROM) 702, erasable programmable read-only memory (EPROM), flash memory, optical fiber, portable compact disc read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this application, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying a computer-readable computer program. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media can also be any computer-readable medium other than computer-readable storage media, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The computer program contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to wireless, wired, etc., or any suitable combination thereof.

[0103] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. Each block in the flowchart or block diagram may represent a module, program segment, or portion of code, which contains one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, or they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block or combination of blocks in the block diagram or flowchart may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0104] The units described in the embodiments of this application can be implemented in software or hardware, and the described units can also be located in a processor. The names of these units do not necessarily limit the specific unit itself.

[0105] Another aspect of this application provides a computer-readable storage medium storing a computer program that, when executed by a computer's processor, causes the computer to perform the object pose estimation method as described above. This computer-readable storage medium may be included in the memory described in the above embodiments, or it may exist independently and not incorporated into that memory.

[0106] Another aspect of this application provides a computer program product or computer program including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the object pose estimation methods provided in the various embodiments described above.

[0107] The above embodiments are merely illustrative of the principles and effects of this application and are not intended to limit this application. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of this application. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in this application should still be covered by the claims of this application.

Claims

1. A method of object pose estimation, the method comprising: include: Obtain the RGB image to be processed and its point cloud representation; The RGB features of the RGB image and the point cloud features of the point cloud representation are extracted respectively, and the RGB features and the point cloud features are fused to obtain the first fused feature; The current time step of the diffusion model is encoded to generate a time step feature, and the time step feature is fused with the first fusion feature to obtain a second fusion feature; The second fused feature is input into the pre-trained diffusion model. During the backward diffusion process of the diffusion model, the second fused feature is used as the guiding condition input. The noise in the noisy pose is predicted by the denoising network of the diffusion model, and the noise in the noisy pose is removed to reconstruct the target pose. The noisy pose is obtained by adding Gaussian noise to the real target pose during the forward diffusion process of the diffusion model.

2. The object pose estimation method according to claim 1, characterized in that, The RGB features and the point cloud features are fused to obtain a first fused feature, including: The RGB features and the point cloud features are concatenated to obtain the concatenated features; The spliced ​​features are sequentially processed through convolution and normalization to obtain the gating weights; The first fused feature is obtained by weighting and summing the RGB features and the point cloud features using the gating weights.

3. The object pose estimation method according to claim 1, characterized in that, The time-step features are fused with the first fusion feature to obtain the second fusion feature, which includes: Obtain a random mask variable, and process the first fused feature based on the random mask variable; The first fusion feature, after being processed by the random mask variable, is added element-wise to the time step feature to generate the second fusion feature; wherein, the probability that the first fusion feature is a null value is the first probability, and the probability that the first fusion feature is a true condition is the second probability, and the first probability and the second probability are complementary.

4. The object pose estimation method according to claim 3, characterized in that, Predicting noise in noisy poses using the denoising network of the aforementioned diffusion model includes: Perform one prediction based on the true conditions and one prediction based on null values ​​respectively to obtain the conditional prediction noise corresponding to the true conditions and the unconditional prediction noise corresponding to the null values. The conditional prediction noise and the unconditional prediction noise are weighted using a preset guiding weight hyperparameter to obtain the noise in the noisy posture.

5. The object pose estimation method of claim 1, wherein, Predicting noise in noisy poses using the denoising network of the aforementioned diffusion model includes: The noisy input of the diffusion model is encoded using a multilayer perceptron to obtain pose features; The posture features are concatenated with the guiding condition input to obtain the network input features of the denoising network; The network input features are processed by the denoising network to obtain effective features; The effective features are added element by element to the network input features; Based on the features obtained by element-wise addition and the multilayer perceptron, the noise in the noisy pose is obtained.

6. The object pose estimation method of claim 5, wherein, The denoising network includes at least one backdiffusion block, which includes residual units and self-attention units; the network input features are processed through the denoising network, including: The residual unit performs feature transformation on the network input features; The global features are obtained by capturing the global dependencies between different positions in the features after feature transformation through the self-attention unit.

7. The object pose estimation method of claim 1, wherein, Obtaining the point cloud representation of the RGB image includes: A coarse depth estimation is performed on the RGB image to obtain a depth map; The RGB image is segmented into instances to obtain a binary mask; The depth map is cropped using the binary mask, and the cropped depth map is converted into a point cloud representation through back projection.

8. An object pose estimation apparatus characterized by comprising: include: The data acquisition module is used to acquire the RGB image to be processed and the point cloud representation of the RGB image; The first feature fusion module is used to extract the RGB features of the RGB image and the point cloud features of the point cloud representation, and fuse the RGB features and the point cloud features to obtain the first fused feature; The second feature fusion module is used to encode the current time step of the diffusion model to generate time step features, and fuse the time step features with the first fusion feature to obtain the second fusion feature; The pose reconstruction module is used to input the second fused feature into a pre-trained diffusion model. During the backward diffusion process of the diffusion model, the second fused feature is used as a guiding condition input. The noise in the noisy pose is predicted by the denoising network of the diffusion model, and the noise in the noisy pose is removed to reconstruct the target pose. The noisy pose is obtained by adding Gaussian noise to the real target pose during the forward diffusion process of the diffusion model.

9. An electronic device, comprising: include: One or more processors; A storage device for storing one or more programs, which, when executed by the one or more processors, cause the electronic device to implement the object pose estimation method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by the computer's processor, causes the computer to perform the object pose estimation method according to any one of claims 1 to 7.

Images